This article was downloaded by: [Institute of Remote Sensing Application]On: 05 March 2014, At: 17:25Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Remote Sensing LettersPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/trsl20

Automatic land-cover update approachintegrating iterative training sampleselection and a Markov Random FieldmodelKun Jiaa, Shunlin Liangab, Xiangqin Weic, Lei Zhangc, Yunjun Yaoa

& Shuai Gaoc

a State Key Laboratory of Remote Sensing Science, and Collegeof Global Change and Earth System Science, Beijing NormalUniversity, Beijing 100875, Chinab Department of Geographical Sciences, University of Maryland,College Park, MD 20742, USAc Institute of Remote Sensing and Digital Earth, Chinese Academyof Sciences, Beijing 100101, ChinaPublished online: 05 Mar 2014.

To cite this article: Kun Jia, Shunlin Liang, Xiangqin Wei, Lei Zhang, Yunjun Yao & Shuai Gao (2014)Automatic land-cover update approach integrating iterative training sample selection and a MarkovRandom Field model, Remote Sensing Letters, 5:2, 148-156

To link to this article: http://dx.doi.org/10.1080/2150704X.2014.889862

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the“Content”) contained in the publications on our platform. However, Taylor & Francis,our agents, and our licensors make no representations or warranties whatsoever as tothe accuracy, completeness, or suitability for any purpose of the Content. Any opinionsand views expressed in this publication are the opinions and views of the authors,and are not the views of or endorsed by Taylor & Francis. The accuracy of the Contentshould not be relied upon and should be independently verified with primary sourcesof information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever orhowsoever caused arising directly or indirectly in connection with, in relation to or arisingout of the use of the Content.

This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,

systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014

Automatic land-cover update approach integrating iterative trainingsample selection and a Markov Random Field model

Kun Jiaa, Shunlin Lianga,b, Xiangqin Weic*, Lei Zhangc, Yunjun Yaoa, and Shuai Gaoc

aState Key Laboratory of Remote Sensing Science, and College of Global Change and Earth SystemScience, Beijing Normal University, Beijing 100875, China; bDepartment of Geographical Sciences,University of Maryland, College Park, MD 20742, USA; cInstitute of Remote Sensing and Digital

Earth, Chinese Academy of Sciences, Beijing 100101, China

(Received 7 December 2013; in final form 24 January 2014)

Land-cover updating from remote-sensing data is an effective means of obtainingtimely land-cover information. An automatic approach integrating iterative trainingsample selection (ITSS) and a Markov Random Field (MRF) model is proposed in thisstudy to overcome the land-cover update problem when no previous remote-sensingdata corresponding to the land-cover data are available. A case study in the Beijingregion indicates that ITSS can effectively select reliable training samples based onhistorical land-cover data and that ITSS with MRF can achieve satisfactory land-coverupdate results (overall classification accuracy: 83.1%). The MRF model can effectivelyreduce salt-and-pepper noise and improve overall accuracy by approximately 6%. Theproposed approach is completely unsupervised and has no strict requirements fornewly acquired remote-sensing data for land-cover updating.

1. Introduction

Global and regional land-cover status and changes are fundamentally important for climateand environmental change studies (Turner, Lambin, and Reenberg 2007; Jin et al. 2013).Moreimportantly, accurate and timely land-cover information is an essential factor for improvingthe performance of ecosystem, hydrologic and atmospheric models (Bounoua et al. 2002; Junget al. 2006; Miller, Guertin, and Goodrich 2007). Due to natural causes and human activities,land-cover is changing throughout the world. Therefore, it is important to quantify andmonitor these land-cover changes to support global and environmental changes studies.

Remote sensing is an effective means for land-cover monitoring with its ability toprovide quickly broad, precise, impartial and easily available information on the landsurface (Hansen et al. 2000; Liu et al. 2003). Many global and regional land-cover datasets have been derived from remote-sensing data (Hansen et al. 2000; Friedl et al. 2002;Liu et al. 2010; Gong et al. 2013), but these data sets use only data acquired during one orseveral years and represent land-cover characteristics for a specific period, without long-term change information. Land-cover update using timely remote-sensing data to supportglobal change studies is one feasible strategy. Furthermore, the large and growing satellitedata archives make it possible to achieve this requirement.

Traditional land-cover updating approaches determine and classify changed areas basedon co-registered multi-temporal remote-sensing data analysis (Yang et al. 2003; Chen et al.2012; Jin et al. 2013). Spectral-based change-detection methods are commonly used for

*Corresponding author. Email: weixq@radi.ac.cn

Remote Sensing Letters, 2014Vol. 5, No. 2, 148–156, http://dx.doi.org/10.1080/2150704X.2014.889862

© 2014 Taylor & Francis

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014

land-cover change detection; for example, Change Vector Analysis has been proposed toupdate the 2001 national land-cover database (Xian, Homer, and Fry 2009). Spectral-basedchange-detection methods can produce an accurate land-cover map, but have strict require-ments for remote-sensing data, specifically that two sets of data acquired in different yearsshould come from the same phenological period. Furthermore, nearly all current methodsneed historical remote-sensing data corresponding to the historical land-cover data forchange detection. If only historical land-cover data are available and the correspondingremote-sensing data are missing, or if the land-cover data are not produced using remote-sensing data, change-detection-based land-cover updating methods will lose efficacy.Therefore, developing robust, efficient and accurate automated or semi-automated methodsfor cost-effective update of land-cover maps in this situation is a challenging work.

The primary goal of this research is to develop and evaluate an automatic land-coverupdating approach using remote-sensing data in situations where no historical remote-sensingdata are available. First, an iterated training samples selection (ITSS) procedure is used toselect training samples automatically based on historical land-cover data. Then the samples areused to classify newly acquired remote-sensing data, and the map is refined based oncomparison of the classification result with that of the last iteration until the difference betweentwo successive classification maps achieves a satisfactory consistency. In addition, a MarkovRandom Field (MRF) model is integrated into the classification procedure to reduce salt-and-pepper noise which is usually apparent in pixel-based classification of remote-sensing data.

2. Study area and data

2.1. Study area

A rectangular subset region of Beijing is selected to validate the presented approach. Thestudy area is located between latitude 39º58′N and 40º9′N and longitude 116º11′E and116º30′E, with an area of about 500 km2. Beijing is located in the northern part of theNorth China Plain and belongs to the temperate climatic zone. The average annualtemperature and precipitation is about 12ºC and 664 mm, respectively. The numerousland-cover types in this area, including forest, cropland, built-up areas and water, havemade land-cover updating in this region a representative of the difficulties encountered.Furthermore, Beijing has undergone rapid urbanization and economic growth in recentyears, especially after the 29th Olympic Games. Therefore, a land-cover map updateexperiment in this region is a suitable choice.

2.2. Data and processing

A Landsat 8 Operational Land Imager (OLI) data set (path/row: 123/32) of the study areaon 12 May 2013 was downloaded from the United State Geological Survey website(http://glovis.usgs.gov/). The OLI data was not affected by cloud, and the quality of themultispectral data was good. OLI data processing mainly included radiance calibrationand subset selection. Radiance calibration was performed to convert the digital numbers(DN) value to surface spectral reflectance. Atmospheric correction was done usingFLAASH tools provided by ENVI version 5.0 (Exelis, Visual Information Solutions,Boulder, CO, USA). Subset selection was done to extract the data covering the study areafor land-cover map updating.

The finer-resolution global land-cover data set produced by Gong et al. (2013) wasselected as the historical land-cover, which was based on classification of Landsat TM/

Automatic land cover update 149

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014

ETM+ data using manually selected training samples and a support vector machineclassifier, and the overall classification accuracy was 71.5% based on the validationsamples. The Landsat data used for classification in this land-cover map was acquiredon 22 September 2009. According to the class type distribution in the subset region, fiveland-cover types were selected as the final classification types, including water, crops,forest, impervious surfaces and bare land.

3. Method

3.1. Classification method

The Maximum-likelihood classifier (MLC) is selected for land-cover map updating in thisstudy because MLC offers high calculation speed with large training samples, unlike somenon-parametric classification methods. MLC classifier assumes that a hyper-ellipsoiddecision volume can be used to approximate the shape of the data clusters. For a givenunknown pixel, described by a vector of features, the probability of membership in eachclass is calculated using the mean feature vectors of the classes, the covariance matrix andthe prior probability (Jia et al. 2011). The unknown pixel is considered to belong to theclass with the maximum probability of membership.

3.2. Markov random field model

The MRF model is usually used to reduce salt-and-pepper noise in pixel-based remote-sensing data classification (Bruzzone and Prieto 2000; Chen et al. 2012). In the MRFmodel, the probability of a pixel belonging to a given class (Cl) is determined both by thespectral information and by its neighbouring pixels.

PðClði; jÞÞ ¼ 1

Zexp½�UðClði; jÞÞ� (1)

UðClði; jÞÞ ¼ Ucontext ðClði; jÞÞ þ Uspectrum ðClði; jÞÞ (2)

where Z is a normalizing factor and U, Ucontext and Uspectrum are the total energy, energy ofthe context and energy of the spectrum. Uspectrum is the log function of the posteriorprobability

Uspectrum ðClði; jÞÞ ¼ � ln ðPspectrum ðClði; jÞÞÞ (3)

where Pspectrum is the posterior probability derived from the MLC classifier. The energy ofthe context is calculated based on the labels of the neighbouring pixels (Bruzzone andPrieto 2000).

Ucontext ðClði; jÞÞ ¼ Ucontext ðClði; jÞ=fClðg; hÞ; ðg; hÞ 2 Nði; jÞgÞ¼

Xðg;hÞ2Nði;jÞ

βδkðClði; jÞ;Clðg; hÞÞ (4)

where β is a constant set to 1.6 in this study, N(i,j) is a set of neighbouring pixels set to asecond-order neighbourhood with N(i, j) = {(i ± 1, j), (i, j ± 1), (i + 1, j ± 1), (i-1, j ± 1)},and δk is expressed as:

150 K. Jia et al.

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014

δkðClði; jÞ;Clðg; hÞÞ ¼ �1; if Clði; jÞ ¼ Clðg; hÞ0; if Clði; jÞ�Clðg; hÞ

�(5)

Land-cover mapping based on the MRF model classify an image by minimizing the totalenergy function in Equation (1) using an optimization algorithm. In this study, the commonlyused iterated conditional modes (ICM) approach is used to optimize the energy functionbecause ICM has acceptable efficiency and accuracy in the classification of remotely sensedimages (Bruzzone and Prieto 2000; Liu et al. 2008). For all changed pixels,Cl(i, j) is updated tothe class that minimizes the total energy function in Equation (1), and this process is repeateduntil convergence is reached. Finally, isolated pixels are more likely to be replaced by theirneighbouring class type, thereby improving the spatial consistency of the classification map.

3.3. Iterative training sample selection

Automatic training sample selection is performed under the hypothesis that land-coverchanges occur only in small areas, and then the training samples are selected in theunchanged areas based on historical land-cover data. An iterative procedure has beendeveloped to refine the unchanged area to select the training samples, which includes thefollowing steps: (1) all pixels in the historical land-cover map are initially selected as thetraining sample; (2) the training samples are used to train the MLC classifier for newlyacquired remote-sensing data classification and to calculate the posterior probabilitywhich is used for the MRF model; (3) the classification result is refined using the MRFmodel; (4) the refined classification result is compared with that from the last iteration todetect the changed and unchanged pixels, and the changed pixels are removed from thetraining samples to obtain the refined pixels; (5) the consistency rate, which is defined asthe ratio of the number of unchanged pixels to the total number of pixels, is calculated. Ifthe consistency rate reaches 99% compared to the map produced at the previous iteration,the iteration stops, and the classification result of the last iteration is confirmed as the finalupdated land-cover map. Otherwise, the refined training samples are used to classify theremote-sensing data; (6) steps (2)–(5) are repeated until the consistency rate reaches 99%compared to the map produced at the previous iteration.

This iterative procedure is expected to refine the training samples to improve classi-fication accuracy by selecting the unchanged areas at every iteration step. A higherconsistency rate indicates that changes in the training samples have less influence onremote-sensing data classification and that the final training samples have been refined toclassify the remote-sensing data. Consequently, the final updated land-cover map isobtained by classifying the remote-sensing data using the final refined training samples.

3.4. Classification accuracy assessment

To validate the performance of land-cover map updating by the ITSS and MRFapproaches using Landsat OLI data, the classification results were assessed by visualobservation and quantitative classification accuracy indicators. Randomly selected samplepixels were used to assess land-cover classification accuracy quantitatively. Each valida-tion sample was identified by visual interpretation with the help of GoogleEarth and theresearchers’ knowledge and experience. The final numbers of sample pixels for classifica-tion accuracy estimation were 47 pixels for water, 216 pixels for crops, 255 pixels forforest, 586 pixels for impervious surfaces and 85 pixels for bare land. The overall

Automatic land cover update 151

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014

classification accuracy, producer’s accuracy and user’s accuracy were then estimated forquantitative classification performance analysis (Congalton and Green 1999; Tso andMather 2001; Foody 2009, 2002).

4. Results

4.1. Land-cover updating using ITSS and the MRF model

Figure 1 shows the relationship between consistency rate and number of iterations. ITSSwith or without the MRF model algorithm rapidly reaches the 99% consistency rate by thefifth or sixth iteration. The results indicate that the ITSS process can effectively refine thetraining samples and achieve a stable classification result. Furthermore, ITSS with the MRFmodel reaches the 99% consistency rate more quickly than without the MRF model, whichsuggests that the MRF model has a significant effect on the land-cover updating approach.

The land-cover updating results using ITSS with and without the MRF model areshown in Figure 2. From the visual perspective, each class type could be identifiedeffectively using ITSS with or without the MRF model based on expert knowledge,

Iterration number

0.5

Con

sist

ence

Rat

e

0.6

0.7

0.8

0.9

1.0

0 1 2 3 4 5 6 7

Algorithm with MRFAlgorithm without MRF

Figure 1. Relationship between consistency rate and number of iterations.

Legend

Bareland

Water

ImperviousForestCrop

Figure 2. Land-cover updating using the ITSS with (left) and without (right) the MRF model.

152 K. Jia et al.

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014

indicating satisfactory land-cover updating results using the proposed approach. Forestwas distributed mainly in the western mountain regions of Beijing and the BeijingOlympic Forest Park region. Crops and bare land were distributed mainly in the northernplain regions of the study area, while impervious surfaces were distributed mainly inurban regions. The main difference in the land-cover updating results using ITSS with andwithout the MRF model was that ITSS with the MRF model could effectively reduce salt-and-pepper noise in the classification map and achieve a smoother result, although somethin linear objects could also be mistakenly removed. In general, the distribution of land-cover types in the updated map was consistent with the actual situation, and ITSS with theMRF model achieved smoother land-cover classification results.

4.2. Classification accuracy of land-cover updating

The confusion matrixes of land-cover updating using ITSS with and without the MRFmodel are shown in Table 1. Satisfactory land-cover updating results were achieved usingITSS both with and without the MRF model. The overall performance of ITSS with theMRF model approach (overall accuracy 83.1%) was better than that without the MRFmodel (overall accuracy 77.3%). In other words, the MRF model improved the overallclassification accuracy by approximately 6%. These results suggested that the MRF modelcould significantly improve classification accuracy when using the ITSS approach.

Bare land had the lowest user and producer accuracy and showed the maximumconfusion with impervious surfaces and crops (Table 1). Other class types all showedbetter separation from each other, with higher user and producer accuracy. The confusionof bare land and impervious surfaces was caused mainly by the similar spectral features ofthese two class types, or perhaps by mislabelling in the historical data set. In addition,bare land did not always remain bare throughout the year, because bare land could later beplanted with crops. The different phenological periods of the acquired OLI data and theremote-sensing data used for historical land-cover classification brought about errors intraining sample selection and resulted in misclassification of bare land and crops. Overall,

Table 1. Confusion matrixes for land-cover update using the ITSS with and without MRF model.

Mappedclass

Ground truth (pixels)

User accuracy(%)

Produceraccuracy (%)Water Crop Forest Impervious

Bareland Total

ITSS with MRF model

Water 43 0 1 4 0 48 89.6 91.5Crop 0 199 28 48 25 300 66.3 92.1Forest 2 11 213 0 0 226 94.3 83.5Impervious 2 5 13 486 13 519 93.6 82.9Bare land 0 1 0 48 47 96 48.9 55.3

ITSS without MRF model

Water 39 1 4 4 0 48 81.3 83.0Crop 0 199 47 66 22 334 59.6 92.1Forest 4 6 195 4 0 209 93.3 76.5Impervious 4 8 9 429 6 456 94.1 73.2Bare land 0 2 0 83 57 142 40.1 67.1

Total pixels 47 216 255 586 85 1189

Automatic land cover update 153

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014

the integration of ITSS with the MRF model showed satisfactory performance in land-cover updating when no previous remote-sensing data corresponding to the historicalland-cover map were available.

5. Discussion

The proposed approach is completely automatic without any human interaction, which over-comes the problems of time cost and labour intensiveness in traditional supervised classifica-tion depending on manual training sample selection. The automatic approach also avoidstraining sample selection error arising from different interpreters. The ITSS process can selectreliable training samples and achieve satisfactory land-cover update results, while the iterativeprocess converges in no more than six iterations, keeping computational cost low.

Traditional land-cover update methods usually detect changed areas first throughcomparison between newly acquired and historical remote-sensing data and then re-classifythe changed areas to complete the update process. However, sometimes the historicalremote-sensing data corresponding to the historical land-cover data are missing, or thehistorical land-cover data were not produced by classifying remote-sensing data, forexample, from ground survey mapping. The proposed approach has been developed pre-cisely to overcome the problem of missing historical remote-sensing data corresponding tohistorical land-cover data. This approach can extend the land-cover data source selectionrange and use older land-cover data which were not produced using remote-sensing data.

The proposed approach has no strict requirements for newly acquired remote-sensingdata. The only condition is that the newly acquired remote sensing and historical land-cover data must have the same spatial resolution. Compared to some change-detectionmethods which need two sets of remote-sensing data which were acquired at the samephenological period in different years or from the same sensor, the proposed approach cangreatly extend the remote-sensing data-source selection range (Chen et al. 2003; Xian,Homer, and Fry 2009). The approach thus overcomes the difficulty in acquiring remote-sensing data from the exactly same phenological period, free of cloud influence.Furthermore, spectral attenuation caused by an aging satellite sensor may lead to errorsin change-detection processes using spectral change analysis, but the proposed approachavoids the spectral change-detection process. Therefore, much more newly acquiredsatellite remote-sensing data can be used to update historical land-cover data.

The proposed approach has some potential limitations. First, it assumes small land-cover changes between the newly acquired data period and the historical data period. Iflarge land-cover changes have occurred, the initial selection of training samples in theITSS process will select large incorrect samples, leading to misclassification in largeareas and loss of efficacy in the training sample refinement process. This phenomenoncan also occur when historical land-cover data are of poor quality, because error propaga-tion is complex and small errors in training samples may be a source of major MIS-interpretations as found by Foody (2010, 2013). In addition, if one of the land-covertypes represented by a very small number of pixels in the historical land-cover data, thisclass type may have low classification accuracy in the updated land-cover map becausethe small number of pixels may be classified into other types. Moreover, if a new classhas appeared between data acquisitions, the proposed method cannot recognize the newclass type. Furthermore, the MRF model may remove linear or small objects, and there-fore the proposed approach is not suitable for situations where the land surface exhibits ahigh degree of fragmentation or where the land-cover update objective is to examinesmall or linear land-cover objects, such as roads.

154 K. Jia et al.

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014

In this study, MLC is selected as the classification method because it is popular and easilyimplemented, and the MLC classifier is also suitable for large training samples. However, atpresent, manymore advanced non-parametric classification methods have been developed forland-cover classification to improve accuracy, such as support vector machine, neural net-works and design trees (Lu andWeng 2007). Non-parametric classifiers will be investigated toimprove the performance of the proposed approach in future research. In addition, in thisresearch, only one temporal remote-sensing data set is used to update the historical land-covermap. Multi-temporal information, which can be helpful to improve classification accuracy, isnot involved in this study, but will be in future work to improve land-cover update accuracy.

6. Conclusions

This paper presents an automatic approach integrating ITSS and a MRF model for land-cover update in situations where no previous remote-sensing data corresponding to theland-cover map are available. A case study in Beijing using OLI data indicates that theproposed method can achieve satisfactory land-cover updating results, and that the MRFmodel can effectively reduce salt-and-pepper noise and improve classification accuracy.The proposed approach is simple and easy to use for updating land-cover withoutcorresponding historical remote-sensing data. Further research will focus on involvingadvanced non-parametric classifiers in the remote-sensing data classification step andusing multi-temporal remote-sensing data to improve the land-cover update accuracy.

AcknowledgementsThe authors are very grateful for assistance and comments provided by Professor Xingfa Gu ofInstitute of Remote Sensing and Digital Earth, Chinese Academy of Sciences.

FundingThis study was partially supported by the Fundamental Research Funds for the Central Universities,and the Science and Technology Subsidies to Leading Scientists by National Administration ofSurveying, Mapping and Geoinformation.

ReferencesBounoua, L., R. Defries, G. J. Collatz, P. Sellers, and H. Khan. 2002. “Effects of Land Cover

Conversion on Surface Climate.” Climatic Change 52: 29–64. doi:10.1023/A:1013051420309Bruzzone, L., and D. F. Prieto. 2000. “Automatic Analysis of the Difference Image for

Unsupervised Change Detection.” IEEE Transactions on Geoscience and Remote Sensing 38:1171–1182. doi:10.1109/36.843009

Chen, J., P. Gong, C. Y. He, R. L. Pu, and P. J. Shi. 2003. “Land-Use/Land-Cover Change DetectionUsing Improved Change-Vector Analysis.” Photogrammetric Engineering and Remote Sensing69: 369–379. doi:10.14358/PERS.69.4.369

Chen, X. H., J. Chen, Y. S. Shi, and Y. Yamaguchi. 2012. “An Automated Approach for UpdatingLand Cover Maps Based on Integrated Change Detection and Classification Methods.” ISPRSJournal of Photogrammetry and Remote Sensing 71: 86–95. doi:10.1016/j.isprsjprs.2012.05.006

Congalton, R. G., and K. Green. 1999. Assessing the Accuracy of Remotely Sensed Data: Principlesand Practices. Florida: Lewis.

Foody, G. M. 2002. “Status of Land Cover Classification Accuracy Assessment.” Remote Sensing ofEnvironment 80: 185–201. doi:10.1016/S0034-4257(01)00295-4

Automatic land cover update 155

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014

Foody, G. M. 2009. “Classification Accuracy Comparison: Hypothesis Tests and the Use ofConfidence Intervals in Evaluations of Difference, Equivalence and Non-inferiority.” RemoteSensing of Environment 113: 1658–1663. doi:10.1016/j.rse.2009.03.014

Foody, G. M. 2010. “Assessing the Accuracy of Land Cover Change with Imperfect Ground ReferenceData.” Remote Sensing of Environment 114: 2271–2285. doi:10.1016/j.rse.2010.05.003

Foody, G. M. 2013. “Ground Reference Data Error and the MIS-estimation of the Area of LandCover Change as a Function of Its Abundance.” Remote Sensing Letters 4: 783–792.doi:10.1080/2150704X.2013.798708

Friedl, M. A., D. K. Mciver, J. C. F. Hodges, X. Y. Zhang, D. Muchoney, A. H. Strahler, C. E.Woodcock, S. Gopal, A. Schneider, A. Cooper, A. Baccini, F. Gao, and C. Schaaf. 2002.“Global Land Cover Mapping from MODIS: Algorithms and Early Results.” Remote Sensing ofEnvironment 83: 287–302. doi:10.1016/S0034-4257(02)00078-0

Gong, P., J.Wang, L. Yu, Y. C. Zhao, Y. Y. Zhao, L. Liang, Z. G. Niu, X.M. Huang, H. H. Fu, S. Liu, C. C.Li,X.Y.Li,W.Fu,C.X.Liu,Y.Xu,X.Y.Wang,Q.Cheng,L.Y.Hu,W.B.Yao,H.Zhang, P.Zhu,Z.Y.Zhao, H. Y. Zhang, Y.M. Zheng, L. Y. Ji, Y.W. Zhang, H. Chen, A. Yan, J. H. Guo, L.Wang, X. J. Liu,T.T. Shi,M.H.Zhu,Y.L.Chen,G.W.Yang, P.Tang,B.Xu,C.Giri,N.Clinton,Z.L. Zhu, J.Chen, andJ. Chen. 2013. “Finer Resolution Observation and Monitoring of Global Land Cover: First MappingResults with Landsat TM and ETM+ data.” International Journal of Remote Sensing 34: 2607–2654.

Hansen, M. C., R. S. Defries, J. R. G. Townshend, and R. Sohlberg. 2000. “Global Land CoverClassification at 1 km Spatial Resolution Using a Classification Tree Approach.” InternationalJournal of Remote Sensing 21: 1331–1364. doi:10.1080/014311600210209

Jia, K., B. F. Wu, Y. C. Tian, Y. Zeng, and Q. Z. Li. 2011. “Vegetation Classification Method withBiochemical Composition Estimated from Remote Sensing Data.“ International Journal ofRemote Sensing 32: 9307–9325. doi:10.1080/01431161.2011.554454

Jin, S. M., L. M. Yang, P. Danielson, C. Homer, J. Fry, and G. Xian. 2013. “A ComprehensiveChange Detection Method for Updating the National Land Cover Database to Circa 2011.”Remote Sensing of Environment 132: 159–175. doi:10.1016/j.rse.2013.01.012

Jung, M., K. Henkel, M. Herold, and G. Churkina. 2006. “Exploiting Synergies of Global LandCover Products for Carbon Cycle Modeling.” Remote Sensing of Environment 101: 534–553.doi:10.1016/j.rse.2006.01.020

Liu, D. S., K. Song, J. R. G. Townshend, and P. Gong. 2008. “Using Local Transition ProbabilityModels in Markov Random Fields for Forest Change Detection.” Remote Sensing ofEnvironment 112: 2222–2231. doi:10.1016/j.rse.2007.10.002

Liu, J. Y., Z. X. Zhang, X. L. Xu, W. H. Kuang, W. C. Zhou, S. W. Zhang, R. D. Li, C. Z. Yan, D. S.Yu, S. X. Wu, and N. Jiang. 2010. “Spatial Patterns and Driving Forces of Land Use Change inChina During the Early 21st Century.” Journal of Geographical Sciences 20: 483–494.doi:10.1007/s11442-010-0483-4

Liu, J. Y., D. F. Zhuang, D. Luo, and X. Xiao. 2003. “Land-Cover Classification of China:Integrated Analysis of AVHRR Imagery and Geophysical Data.” International Journal ofRemote Sensing 24: 2485–2500. doi:10.1080/01431160110115582

Lu, D., and Q. Weng. 2007. “A Survey of Image Classification Methods and Techniques forImproving Classification Performance.” International Journal of Remote Sensing 28: 823–870. doi:10.1080/01431160600746456

Miller, S. N., D. P. Guertin, and D. C. Goodrich. 2007. “Hydrologic Modeling Uncertainty Resultingfrom Land Cover Misclassification.” Journal of the American Water Resources Association 43:1065–1075. doi:10.1111/j.1752-1688.2007.00088.x

Tso, B., and P. M. Mather. 2001. Classification Methods for Remotely Sensed Data. London: Taylorand Francis. doi:10.4324/9780203303566

Turner, B. L., E. F. Lambin, and A. Reenberg. 2007. “The Emergence of Land Change Science forGlobal Environmental Change and Sustainability.” Proceedings of the National Academy ofSciences of the United States of America 104: 20666–20671. doi:10.1073/pnas.0704119104

Xian, G., C. Homer, and J. Fry. 2009. “Updating the 2001 National Land Cover Database LandCover Classification to 2006 by Using Landsat Imagery Change Detection Methods.” RemoteSensing of Environment 113: 1133–1147. doi:10.1016/j.rse.2009.02.004

Yang, L. M., G. Xian, J. M. Klaver, and B. Deal. 2003. “Urban Land-Cover Change Detectionthrough Sub-Pixel Imperviousness Mapping Using Remotely Sensed Data.” PhotogrammetricEngineering and Remote Sensing 69: 1003–1010. doi:10.14358/PERS.69.9.1003

156 K. Jia et al.

Dow

nloa

ded

by [

Inst

itute

of

Rem

ote

Sens

ing

App

licat

ion]

at 1

7:25

05

Mar

ch 2

014